A simple five-part framework
What do you want accomplished?
What background does the AI need?
What must it do or avoid?
What should the final result look like?
5. Evidence
Tell the AI what should be verified, cited, tested or checked before the work is considered complete.
Conversation or master prompt?
Conversation is best for exploration, learning and refinement. Master instructions are useful for recurring brand, safety, approval and workflow rules.
Give the model enough context to succeed
Useful context can include the audience, purpose, source material, previous decisions, definitions and constraints that change what a good answer looks like. More context is not automatically better: irrelevant history can distract from the current task. Supply the smallest set of facts that materially changes the result.
For recurring work, keep stable preferences and governance rules separate from the immediate request. Then describe today's objective plainly. This makes it easier to update one part without rewriting everything and reduces contradictions between old and new instructions.
Use iteration as part of the workflow
A strong first instruction should make the direction clear, but complex work often improves through review. Ask the AI to identify missing evidence, test assumptions, compare alternatives or revise against explicit criteria. When a task matters, verify the result using authoritative sources, calculations, tests or human review rather than treating fluent wording as proof of correctness.